I’m not sure I share your certainty on that. It’s possibly true, it’s also possible that the recognizer just runs with a slower clock, but I think even if not it would still be a relatively small amount of memory - it’s not like it is looking at IQ samples for a whole packet, just a symbol and maybe half of each adjacent one.
That said, if there actually is a hardware difference of the actual silicon I’d join you in considering performance binnning at test to be a likely suspect.
My main suspect however would remain software misconfiguration. If you have a setup which works with one module and not another, that would perhaps point to something.
It might be worth doing a test with a node-class radio as receiver so that you could read out the frequency error.
I’ve mostly not been using radios in RFM95 format but have one board re-wired to one (sourced I believe also from digikey), you’re making me a little tempted to dig it out and see if it works, but not sure after a lot of focus on the CMWX1ZZABZ that the code of my main project still builds for the MCU connections of the one-off RFM95 lash-up.